Amazon is winding down most of its flagship Nova AI models in a significant strategic shift, concentrating its resources on a single, larger effort to build one genuinely competitive foundation model. The move gives concrete shape to a retreat that until recently looked mostly like closed offices and job cuts. The company has begun deprecating most of its in-house Nova line, including the high-end Premier and Omni models, the Reel video generator, and the Canvas image generator, according to internal communications and reports. Some staff have reportedly referred to this as “KTLO” mode, short for “keep the lights on,” meaning the models will remain supported for existing customers but are no longer a development priority.
This development is the concrete update to a story that first emerged months ago when Amazon shut its AGI Lab and effectively exited a race it never led. At that time, the news centered on a closed lab and layoffs. Now it is clear which products are being sacrificed, and what they are being sacrificed for. Amazon's ambition in artificial intelligence has always been broad, but the company is now making a deliberate choice to narrow its focus in the face of intense competition from OpenAI, Anthropic, and Google.
One big bet instead of many
Resources are shifting to a new effort called Frontier Model Research, or FMR, led by Pieter Abbeel. Abbeel joined Amazon through its 2024 acquisition of the robotics startup Covariant, bringing deep expertise in machine learning and robotics. A new flagship model is expected to debut at Amazon's re:Invent conference this autumn, as reported by Reuters, and it could still carry the Nova name. This new initiative represents a fundamental change in how Amazon approaches foundation model development.
The logic behind this shift is focus. Under previous AI chief Rohit Prasad, Amazon spread itself across text, image, and video models, attempting to compete on multiple fronts simultaneously. That approach led to a dilution of talent and computational resources. Peter DeSantis, who took over the consolidated AI group in December, has concentrated talent and scarce compute on fewer frontier bets. Prasad left at the end of 2025, and AGI Lab founder David Luan departed in February, signaling a clear break from the previous strategy.
Amazon's Nova model family was launched in late 2024, aiming to provide a range of capabilities from text generation to video creation. The lineup included Nova Premier, a high-end text model designed for complex reasoning tasks; Nova Omni, a multimodal model capable of handling text, image, and video inputs; Nova Reel, a video generator; and Nova Canvas, an image generator. Despite the ambitious range, these models struggled to gain significant traction in a market already dominated by established players with massive user bases and strong brand recognition.
One of the key problems was cost. Amazon's systems proved expensive to operate relative to the value they returned, according to former employees and industry analysts. Cheaper rivals piled on the pressure, part of a wider shift toward cut-price models across the AI industry. As open-source and cost-efficient alternatives emerged from companies like Meta, Mistral, and a host of startups, Amazon found it increasingly difficult to justify the enormous expenditure required to maintain a dozen different models at varying levels of quality.
Where Amazon actually wins
Amazon never made Nova a household name the way OpenAI, Anthropic, and Google did with their models. ChatGPT, Claude, and Gemini have become synonymous with generative AI in both the public and enterprise sectors. Nova, by contrast, remained a niche offering known primarily to AWS customers and developers who were already deeply embedded in the Amazon ecosystem. This lack of brand recognition, combined with performance gaps and high costs, made it clear that Amazon could not compete effectively across the entire model spectrum.
So Amazon leaned into what it does best: infrastructure. AWS is the landlord for much of the AI industry, with compute commitments worth $138 billion from OpenAI and more than $100 billion from Anthropic. These massive deals ensure a steady revenue stream regardless of which company ultimately wins the model race. Amazon's custom Trainium chips are now a multibillion-dollar business that Jeff Bezos has called a fourth company pillar, joining AWS, retail, and advertising. The company is betting that the AI boom's biggest payoff will come not from owning the most intelligent model, but from providing the computing power that every model depends on.
An Amazon spokesperson rejected the idea of a retreat. “AI models remain one of the most important things we're working on, and that hasn't changed,” the spokesperson said, adding that the company “continually evolves” its lineup around what customers need. This careful phrasing suggests that Amazon is not abandoning AI research but rather reallocating resources to maximize impact. The company remains committed to serving customers across its sprawling ecosystem, from e-commerce to cloud computing to digital assistants.
What remains of Nova
Not all of Nova is vanishing. Amazon has confirmed that several products will stay active. The company is keeping Nova 2 Lite, a lightweight and cost-efficient text model, and Nova 2 Sonic, which appears to be a specialized text or audio model. The Nova Forge customization service, which allows customers to fine-tune models on their own data, will also continue, as will the Nova Act agent tool, designed for automation and task completion. These remaining products are focused on practical, high-demand use cases rather than experimental frontier research.
The company's San Francisco AGI site, an 80-person research group, has closed, according to GeekWire. That closure marks another step in the consolidation of Amazon's AI efforts. The San Francisco office was reportedly focused on longer-term research questions about artificial general intelligence, but those ambitions have now been folded into the broader Frontier Model Research initiative. Employees at the site were offered positions elsewhere in the company or layoff packages, reflecting the human cost of this strategic pivot.
The decisions also highlight a broader trend in the AI industry. Companies are increasingly recognizing that building and maintaining a wide portfolio of models is financially unsustainable, even for a tech giant like Amazon. The cost of training large-scale models, securing data, and keeping up with rapid advances requires deep pockets and intense specialization. As a result, several major companies are consolidating their AI efforts around fewer, more ambitious projects rather than spreading resources thin across numerous smaller models.
The competitive landscape
The AI model market has become brutally competitive, with new releases arriving almost weekly. OpenAI's GPT series has become the gold standard for conversational AI, while Anthropic's Claude models are prized for their safety and reasoning capabilities. Google's Gemini family integrates deeply with the company's search and productivity tools, giving it a built-in distribution channel. Amazon, by contrast, lacks a comparable consumer-facing AI product, and its enterprise-focused approach has not generated the same level of excitement or adoption.
Amazon's decision to focus on a single frontier model is a recognition of these realities. The company appears to be following a playbook similar to that of some competitors: concentrate resources on one flagship model that can match or exceed the best available alternatives. The upcoming model from FMR, expected to debut at re:Invent, will be a test of whether Amazon can translate its massive infrastructure advantages into a genuinely competitive foundation model.
Pieter Abbeel is a well-known figure in the AI research community, with a background that includes founding Covariant, a robotics company whose technology was acquired by Amazon in 2024. He is also a professor at UC Berkeley and has been a leading voice in robotics and machine learning for years. His appointment as the head of FMR signals Amazon's desire to bring in established research talent to lead its most important AI initiative. Under his leadership, the team is reportedly focused on developing a model that is not only technically advanced but also cost-effective to serve at scale.
However, the road ahead is far from easy. The compute and data requirements for building a frontier model are staggering, and Amazon's competitors have substantial head starts. Open AI reportedly spends billions of dollars annually on training and inference, and Anthropic and Google have similar scale. Amazon's advantage lies in its vertically integrated stack, including Trainium chips, AWS infrastructure, and a massive distribution network through AWS and its enterprise sales force.
The harder question is the one Nova never answered. Amazon can host everyone else's models and sell the chips underneath them. Whether it can also build a frontier model that developers actively choose over Claude, Gemini, or GPT is what re:Invent will test. The model's performance, pricing, and integrations will all be scrutinized by the AWS customer base, which has become accustomed to a wide range of model options on the platform.
Amazon's move also has implications for the broader AI ecosystem. The company's decisions often shape market dynamics, and a successful frontier model could provide a more competitive marketplace with alternative options to the current leaders. On the other hand, a further retreat could signal that even deep-pocketed companies find it difficult to compete with the frontrunners. Either way, Amazon's bet on a single frontier model marks a defining moment in its AI journey, one that will unfold in the coming months as the company prepares to unveil its latest creation at re:Invent. The outcome may well determine whether Amazon becomes a true AI leader or remains a powerful but secondary player in the fastest-moving technology race of the decade.
Source: TNW | Amazon News